Juniper execs share network vision

Kim Perdikou and David Yen discuss the networking vendor's convergence offerings and tech trends
  • John Dix (Network World)
  • 08 February, 2010 16:29

Juniper Networks has always been about high performance and, since it straddles the carrier and enterprise markets, has an interesting perspective on where these worlds intersect. Network World Editor in Chief John Dix caught up with Kim Perdikou, EVP and GM of the Infrastructure Products Group, and David Yen, EVP and GM of the Fabric and Switching Technologies Business Group, to discuss converging needs, tech trends and the company's Stratus project, a single-layer network architecture.

What kind of overlap do you see with carrier and enterprise networks today?

Perdikou: There's a lot of discussion around cloud computing, with service providers preparing to be cloud ready and enterprises [trying to figure out] what they can leverage from a service point of view. In addition, because of the cost and the complexity of running enterprise wide-area networks, many companies are starting to outsource that to service providers. But for service providers, the network costs are rising faster than the revenue they make on them.

Why is that?

Much of the service providers' business model for bandwidth is all you can eat for a certain price. And that is in danger of breaking. We believe it will break over the next three to four years. How do they transform their business model in order to charge for the right services at the right time that people are willing to pay for it? Let me give you an example. I want to pay a certain amount for bandwidth to my home, but if I'm in the airport about to get on a plane, if I could press a magic button and have all my e-mail downloaded in a minute, I would be willing to pay five or 10 dollars for that. I don't put that in the same context as, "How much am I paying for bandwidth for a month?" So how do you build a network that you can start to charge where it's relevant, that people are willing to pay for?

Is part of the answer to migrate everything to IP?

The vision of one IP infrastructure for all traffic has been accepted as how we get there. The problem is legacy services or at least legacy access. It's all about reducing the time from concept of a service to marketing it and making money out of it. And to do that, they really need to redo the network to make it simpler, to make it cheaper to run. Even on the billing side. One customer I spoke to was celebrating that they went from 300 to 30 billing systems. So, if you have that kind of complexity, getting to one network infrastructure and one billing system is a huge, huge challenge.But there is agreement that IP on Ethernet is the Nirvana vision?

IP Ethernet is the long-term answer, but it's not the only technology. In some cases they'll just use optical transport and put routing or switching around it.

So are enterprise and carrier networks getting more similar?

We focus on customers with high-performance networking needs. If you look at the enterprise switching market, it's probably $21 billion, and we believe about half of that is addressable as high performance. Why? Because the requirements of the enterprise are moving more to what service providers require -- scale, performance at scale, availability. I said to a financial company, "You know to a service provider a 911 call is life and death" and he said to me, "a financial transaction is life or death." So, when the customer believes that, the requirements are coming together.

Yen: I just came back from Greater China Mainland and Taiwan, and while it is not going to change overnight, the focus is on growing IP, growing Ethernet.The announcement you made recently about the MX 3D was about scaling Ethernet for carriers, right?

Perdikou: Yes, and it's interesting because when the market for carrier Ethernet started to open up, people that had switching for LAN environments said, "Oh, we can use these existing products there." One, we didn't have switching at that time, but two, we didn't believe that anyway. Ethernet is the transport but you have to build in flexibility and programmability, and that's what we did with the MX3D. We believe it's the first universal edge in the industry that truly meets what the universal edge was envisioned to be. If you're going to deliver a lot of video traffic, you will have fewer subscribers with a lot more bandwidth each. If you're going to deliver mobile voice traffic you want a large number of subscribers that use a small amount of bandwidth.

So, the fungability between services and subscribers and bandwidth enables service providers to build networks that are open and flexible with the right level of security and the right level of functionality. And the fact that we have Junos Software in the MX and across the product lines, and in security and in the switching, is another industry first and differentiates us in total cost of ownership in a huge way.

I mean, just looking at the MX3D, the total cost of ownership is at least 40% percent less than most of our competition. And the MX3D will use a tenth of the power of anybody else that delivers 2.6 terabytes at the edge. That is huge for helping to transform the business model for a service provider and to reduce the total cost of ownership in the enterprise.

Page Break

There has been a fair amount of industry discussion about the need to delayer enterprise networks to simplify them. What's your take?

Yen: A three-layer structure is conceptually very helpful, particularly in the data center. You have access, aggregation and the core. But with today's technology, particularly what Juniper is offering, we believe pretty much all data centers -- except a handful of large core operators such as Amazon and Google -- can be supported using a two-tier structure rather than resorting to three tiers.

When you reduce a tier, a layer, by definition you save in cost, you save in power and you have better latency that translates to better performance. So this is why we're winning a number of data center deals -- including the New York Stock Exchange -- with our two-tier approach, the fixed configuration EX3200 and EX4200 in the front, and the modular EX 8200 switches in the back. And very frequently, MX will become part of the equation when the customer looks at edge routing.So you squeeze out the aggregation layer.

Perdikou: You always have the access layer because that's the highest volume, but instead of using a middle layer to reduce the number of ports, if you have sufficient port counts at the core, which is ultimately the fabric providing the switching function, you can subsume the aggregation requirement.

Of course, in order to achieve that, if the total number of ports or lines after the access layer is still huge, you need a bigger structure. But the reason we can always accomplish this in two tiers is, at the access layer we offer software technology we call virtual chassis. This is a way to take multiple access switches -- up to 10 -- and aggregate the collection as if they were one logical switch. That simplifies the management but, more importantly, once they aggregate this way you reduce the number of uplinks.

So, it reduces the number of uplinks and reduces the pressure to have a more complicated higher level structure. Then you bring in the high port count modular switch -- which can serve the combined role of core and aggregation.How big a company can this architecture support?

Yen: That type of structure can cover up to 4,000 or 5,000 servers, pretty much all major businesses.

Perdikou: If you think about the ones that it doesn't fit -- like the online business, what we think of as content service providers -- the two-tier approach with very few exceptions can go from the smallest to the largest.

OK, that's two tier. With the Stratus effort you briefly outlined last year you're talking about moving to a single layer, right? Give us the thumbnail.

Yen: We haven't told the world too much, but last year we publicly disclosed we are undertaking a project called Stratus. Marketing picked the name because stratus is a single layer of flat cloud, and it implies what we are trying to achieve. You can think about it as highly scalable, from supporting several hundred 10Gig E ports to tens of thousands of 10Gig E ports, and all at line rate, so it's not a heavily oversubscribed type of implementation.

It's one architecture, very scalable, any-to-any (any ports can connect to any other ports), it's homogenous, fair (there is no bias of, if you go this way or that way it's faster), and most importantly it is lossless. In the Internet you're allowed to drop a packet upon heavy congestion. But in the data center, for communication between servers or between servers and storage, it's totally unacceptable.

Stratus is also a converged fabric, which means instead of the current practice of using Ethernet for IP traffic, Fibre Channel for storage traffic and Infiniband as a low-latency technology, Stratus will support a converged Ethernet fabric that will support all kinds of data center traffic. And on top of that, Stratus will guarantee a very low, worse-case latency and be competitive to InfiniBand.

If your data center is small, you have 50, 200 servers, then the intranet is nothing but the interconnect among servers and storage. The total cost is low, power consumption is low, latency is relatively low. You don't really care.

But then as your scale grows so does the accumulated latency. And it's only going to get worse with multicore, multithreaded microprocessors that can significantly boost the total throughput a server can drive. And with virtualization you want to run four or eight or 16 virtual machines on one physical machine. Suddenly each server is faithfully driving its gigabyte Ethernet lines -- or 10Gig lines -- near the line rate most of the time and the cost of all of this makes the network a first-class citizen in the data center, just as much a concern as the servers, just as much a concern as the storage.

So, that's what inspired Juniper's Stratus vision. It started in Pradeep Sindhu's organization [Juniper's CTO and founder] and Pradeep has been nurturing and fostering the effort and now it has grown into a full-scale project. We've been working on this in total probably more than two years.

How will it be productized? Is it an answer to Cisco's Nexus stuff?

Perdikou: It's a natural thing to try to compare it because Nexus is Cisco's latest, newest push. There are a few areas we are thinking in common. For example, Nexus is striving toward converged Ethernet within the data center, and in that we are completely aligned. We believe the data center is heading in the direction with unified, converged traffic. If you build a 10-lane highway from downtown to the airport, you want all ten lanes usable by different kinds of vehicles -- whether they are four-door sedans or a pickup truck -- so you can smoothly, flexibly handle the traffic volume.

But Cisco's approach is very incremental. You look at the highly touted Nexus 7000, and pretty much all the sales right now are to replace aging Catalyst 6500s. There is very little real converged traffic being driven on the Nexus line, even though it has been in the market for a couple of years. And you look at the product implementation, it's a very incremental evolution from today's switch. It's nowhere near the fundamental architecture change as in Stratus.

Yen: What Stratus presents to the customer is intended to be a very simple image: a very scalable single data center fabric. Under the hood, Juniper is throwing all the experience we've gained over the last 13 [or] 14 years of cutting-edge switching technology and Internet core routing.

Such an offering takes a lot of effort -- the silicon investment, the hardware investment, multiple layers of software. Now, come back to Juniper and Cisco. If you look at it from the technology capability, Cisco has a lot of talented engineers, so we believe our respectable competitor will have the capability. And if you look at it from the resources perspective -- the ability to invest and create such a product -- they are resourceful enough. But from the business perspective, there's a significant difference between us.

Cisco is already very dominant as a data center intranetworking provider, while Juniper has very little share. So where is Cisco's motivation to say, "Let's introduce a fundamentally new architecture to revamp the whole thing." Moving the money from your left pocket into your right pocket is not going to make you any richer. So, the business incentive for Cisco to significantly invest in a very fast pace of evolution is not there.

Ever since we publicly disclosed Stratus last February we are getting overwhelming interest from all over the world. Everybody wants to call and talk to us. It's similar to -- before I joined Juniper I spent 20 years at Sun Microsystems -- when Sun invented Java in 1995. The phone was ringing off the hook because everyone wanted to know something about Java. And right now, while it will still take some time for Stratus to become available, we are actually enjoying such publicity.

What is the timeframe for Stratus?

Yen: It will take a couple of years. But whenever certain pieces of Stratus technology become available we will offer it on top of today's product.

Stratus inside? (Laughing from group).

Yen: We may not give it the Stratus label but we will apply some pieces of technology when it becomes available.

Is Stratus similar to what Jayshree Ullal and Andy Bechtolsheim are up to at Arista?

Yen: Certainly they have gotten a lot of publicity and they have done a respectable job considering their size. But if you read some of their speeches they say they're pretty much providing the access layer. And when asked how about the core, they actually say the customer will have to go through Cisco or Juniper to complete a whole configuration. They know their limitations.

Shifting gears to computing, what do you make of Cisco's Unified Computing System?

Yen: Having come from the server world, UCS is just the next generation of blade servers. The vendors that reacted to the move were the server companies, IBM, HP, Dell. Cisco literally invaded their market. But instead of joining the very low margin X86 vendors and building conventional X86 servers, they carved out a portion of the network interface, a portion of the storage interface, and then adopted the latest Intel multicore, multithread microprocessor and, with their relationship with Intel, came up with an innovative approach to provide more memory on their blade, which is conducive to the number of virtual machines they can support. Then they laid in the VMware virtualization software and the BMC management software.It's a local scale system integration job. But by doing that integration, obviously, they eliminated some unnecessary hardware/software which reduces the total cost of such an aggregation. And with the reduction of that cost, they could give the customer half of the savings as an incentive and keep half of the saving to raise the margin.

So, it's an interesting move on their part. Unfortunately, they enter territory they don't really know that much about. Getting into servers is not just putting the hardware together, but most seriously, they are hurting their partnerships with those system companies.

Page Break

They claim their architecture will make it easier to marry network policies to mobile virtual machines.

Perdikou: That's marketing speech for a proprietary implementation. Today, every time Intel comes up with a new microprocessor, within a couple of weeks any customer can pick up the phone and get no less than six vendors to come in and bid, all with that latest microprocessor. Now, look at UCS. The next time Intel comes up with a new microprocessor the customer who adopted UCS has to sit there waiting for Cisco to come around to upgrade their system? And when it comes to virtualization, there are multiple proposals right now in standards bodies trying to standardize the virtual machine-to-networking interface, including what kind of a state it has to carry when it migrates from one place to the other, in what format that state be carried, etc. It's a matter of standardization.

Once completed, even when the customer utilizes servers and network equipment from different vendors, the virtual machine should be able to migrate from one server to another server through various kinds of network equipment. It has nothing to do with integrating things together.

UCS is absolutely a lock-in strategy on Cisco's part. Because of my background, customers ask me for my opinion about UCS. And I say, "Well, if you are a small business and you believe for the lifetime of your business you can stay within the scope of UCS, then yes, Cisco has done the pre-integration with virtualization software, management software, did the pretesting, it's ready to go."

But in the long term I don't think it's the right thing to do. The bulk of the industry wants separated computing, server, storage and networking. The customer wants the fungibility, an interchangeable capability.

So, whereas Cisco is bold enough, arrogant enough to say, let's carve out this area and provide the whole thing to the customer, partner or no partners, Juniper focuses on the networking domain, working with all the traditional data center citizens -- the server, storage and appliances vendors -- and even all the on ramp/off ramp vendors -- the NIC card vendors, the HBA vendors. We don't invade their area. That kind of crossing is hurting the fungibility of this whole ecosystem, and it's not really good for the customers.

In closing here then, any final thoughts?

Yen: The IT world is going through a very significant overhaul and it's going to continue for at least another five to 10 years. So, it's a very significant time and very exciting.

Read more about lans and wans in Network World's LANs & WANs section.